AI is not simply going to replace some employees with cheaper software. The more important transformation is the redesign of companies around systems that can learn, execute and improve without waiting for humans.
A few assumptions first.
I don’t know whether LLMs are a path to AGI or ASI in the physical world. I actually doubt that they are, at least by themselves. But they are already good enough to perform a surprisingly large share of purely cognitive work in the modern economy.
The current combination of models, agents, harnesses, loops and orchestration is sufficient to automate a significant part of knowledge work: office tasks, communication, research, information processing and interaction with information systems. And the technology is improving fast.
Models seem to improve roughly every three months. The price for a given level of capability appears to fall roughly by half every six months. Open-source models tend to catch up with the frontier with a lag measured in months, not years.
The naive extrapolation is pretty brutal. If roughly 30% of knowledge-work tasks are already technically automatable, and capability keeps improving while costs keep falling, the share could approach 60% within the next couple of years.
The exact numbers are debatable. The direction is not.
Every company with office work has a problem
For any company with a significant amount of knowledge work, automating at least some 20-30% of that work is becoming technically possible. And, more importantly, inevitable.
Not because CEOs suddenly became excited about AI. Because of competition.
Imagine your closest competitor automates 30-40% of its internal work. Its operating margin improves. It can lower prices, increase service quality, respond faster and reinvest more aggressively. You don’t get to decide whether to participate.
The same process has happened many times before. Once the economics of a new technology become compelling, adoption spreads through an industry until the companies that cannot adapt disappear or are forced onto the new operating model.
The interesting question is therefore not: “Will AI replace jobs?”
It is: “What does a company look like when a large part of its work no longer needs to be performed by humans?” And this is where things get considerably more difficult.
The bottleneck is not technology. It is organizational will.
A company cannot become AI-native by buying an enterprise ChatGPT subscription. It needs to understand its own work.
That means digitizing, recording and building an ontology around essentially everything the organization does: processes, decisions, artifacts, communications, customer interactions, documents and systems. Then, wherever a process can be represented computationally, you need an agent capable of performing it.
That agent needs access to a continuously updated representation of the business: customer messages, contracts, internal communication, product information, competitor announcements, supplier relationships and everything else relevant to the task. Then you need traces.
What did the agent do? Why? What worked? What failed? Where did it need human intervention? And finally, you need evals.
If I had to reduce the entire transformation to two words, they would probably be: automation optimization.
Not implementation. Not prompting. Not “AI adoption.” Optimization.
You continuously measure how well the system performs, identify failures, change the workflow, retrain or replace the agent, and run the loop again.
That is a fundamentally different way of operating a company.
This is a one-year project, not a weekend project
For a company with 1,000+ employees, I would expect this transformation to take at least a year. Probably longer. But speed matters enormously because of the competitive dynamics.
The uncomfortable part is that the transformation may require eliminating, combining or radically redesigning roles that currently exist. And this is where technology stops being the difficult part.
A CEO can approve a new AI platform. It is much harder for a CEO to say: “We probably don’t need half of these roles anymore, and we are going to redesign the company around that fact.” Especially when the CEO is an employee too.
The people responsible for transforming the organization may have the most to lose from the transformation. Middle management can see entire layers of coordination disappearing. Executives can see their own responsibilities being automated. Departments can discover that their historical reason for existence has disappeared.
So organizations have a very strong incentive to move slowly. They can spend millions on AI while changing almost nothing. That may be the most dangerous outcome of all.
The real unit of automation is not the employee
This is, I think, the most important distinction. Replacing “Peter the lawyer” with “Claude the lawyer” doesn’t fundamentally change the organization.
You have simply substituted one worker for another. The interesting transformation happens when you redesign the entire function.
Take marketing. The goal isn’t to replace five marketers with five agents. The goal is to build a system that continuously: finds potential customers → researches them → creates hypotheses → produces campaigns → launches them → measures results → identifies failures → adjusts the strategy → runs the next experiment. And it does this 24/7.
The system becomes better because it learns from its own mistakes. That is a fundamentally different production function. The same logic applies to sales, customer support, legal, finance, research, procurement and eventually almost any other cognitive function.
The competitive advantage won’t come from having access to the same model as everyone else. It will come from having a better system built around that model.
And this creates a second-order effect
Once some companies become dramatically more productive, the effects don’t stop at those companies. Their competitors have to respond. Then their suppliers respond. Then customers change their expectations. Then pricing changes. Then margins change. Then capital flows toward the most productive companies. Then entire industries reorganize around the new cost structure.
This is why I think focusing on “how many jobs will AI eliminate?” misses the interesting part.
The first-order effect is automation.
The second-order effect is competition.
The third-order effect is organizational redesign.
And the fourth-order effect may be an entirely different economy.
The redistribution argument misses something important
There will obviously be enormous political pressure around this transformation.
When productivity gains become unevenly distributed, people will ask governments to slow things down, increase taxation, protect employment and redistribute more of the gains. Some of those policies may be reasonable. But there is a dangerous feedback loop here.
If an economy becomes less productive, growth slows. When growth slows, distribution becomes more politically salient. People perceive the existing distribution as increasingly unfair and demand more redistribution. If that further reduces incentives to invest, hire and build, growth slows again.
Eventually the political system ends up fighting over a shrinking pie. The alternative is not necessarily “let the market do whatever it wants.”
It is much simpler: increase the size of the pie faster than the distribution problem grows.
AI is potentially one of the largest productivity opportunities in decades. Wasting that opportunity because organizations are afraid to change would be an extraordinarily expensive mistake.
There is no moral to this story
At least not a comforting one. Sam Altman is not going to save your company. Elon Musk is not going to save your company. The government probably isn’t going to save your company either.
Every company is ultimately responsible for adapting to its competitive environment. And adaptation will increasingly mean something much more difficult than “implementing AI.” It means redesigning the organization around a new production function.
The companies that win won’t necessarily be the ones with the best models. They will be the ones that figure out how to turn those models into systems that continuously learn, execute and improve. And there may be a new profession hiding in this transition.
Not prompt engineers. Not AI consultants who teach people how to use ChatGPT. People who can design entirely new business processes for agents. People who can look at a traditional organization and ask: “What if we designed this function from scratch, assuming that agents could perform most cognitive tasks, operate continuously, access the company’s entire knowledge base and learn from every failure?”
That sounds like a niche skill today. It probably won’t stay niche for very long.
Competition may force adoption, but the internal principal-agent problem helps explain why spending can rise while redesign stalls. The managers asked to eliminate coordination layers may be evaluating a system that weakens their own authority, so ‘AI adoption’ becomes a safer substitute for changing decision rights. The real unit of transformation may be the decision boundary: who decides, who verifies, and who absorbs failure. What governance structure have you seen overcome that incentive without turning redesign into a permanent committee?